Departament d´Economia Aplicada · Universitat de Barcelona and Institut d’Economia de Barcelona October 2020 Abstract: Using data from the 545 largest European cities, we study

Departament d´Economia Aplicada

Congestion in highways when tolls and railroads matter: Evidence from European cities

Miquel-Àngel Garcia-López / Ilias Pasidis /

Elisabet Viladecans-Marsal

Facultat d´Economia i Empresa

20.11

D

o

c

u

m

e

n

t

d

e

t

r

e

b

a

l

l

Aquest document pertany al Departament d´Economia Aplicada

Data de publicació: Octubre 2020

Departament d´Economia Aplicada

Edifici B

Campus de Bellaterra

08193 Bellaterra

Telèfon: 00 34 935 811 680

E.mail: [email protected]

https://www.uab.cat/departament/economia-aplicada/

Congestion in highways when tolls and railroads matter:Evidence from European cities

Miquel-Àngel Garcia-López∗†Universitat Autònoma de Barcelona and Institut d’Economia de Barcelona

Ilias Pasidis∗‡Institut d’Economia de Barcelona

Elisabet Viladecans-Marsal∗§Universitat de Barcelona and Institut d’Economia de Barcelona

October 2020

Abstract: Using data from the 545 largest European cities, we studywhether the expansion of their highway capacity provides a solutionto the problem of traffic congestion. Our results confirm that in thelong run, and in line with the ’fundamental law of highway conges-tion’, the expansion in cities of lane kilometers causes an increase invehicle traffic that does not solve urban congestion. We disentanglethe increase in traffic due to the increases in coverage and in capacity.We further introduce road pricing and public transit policies in orderto test whether they moderate congestion. Our findings confirm thatthe induced demand is considerably smaller in cities with road pricingschemes, and that congestion decreases with the expansion of publictransportation.

Key words: congestion, highways, Europe, citiesJEL classification: R41, R48

∗Financial support from the Ministerio de Ciencia e Innovación (research projects ECO2016-75941-R and RTI2018-097401-B-I00), Generalitat de Catalunya (research projects 2017SGR796 and 2017SGR1301), and the ”Xarxa deReferència d’R+D+I en Economia Aplicada” is gratefully acknowledged.†Corresponding author. Department of Applied Economics, Universitat Autònoma de Barcelona, Edifici B, Facultat

d’Economia i Empresa, 08193 Cerdanyola del Vallès, Spain (e-mail: [email protected]; phone: +34 93 5814584; website: http://gent.uab.cat/miquelangelgarcialopez).‡John M. Keynes 1-11, 08034 Barcelona, Spain (e-mail: [email protected]; phone: +31 613 087 495; website:

www.iliaspasidis.com).§Department of Economics, Universitat de Barcelona, John M. Keynes 1-11, 08034 Barcelona, Spain (e-mail:

[email protected]; phone: +34 93 403 4646; website: https://elisabetviladecans.wordpress.com/).

[email protected]://gent.uab.cat/miquelangelgarcialopezipasidis@[email protected]://elisabetviladecans.wordpress.com/

1. Introduction

Road congestion remains one of the most pressing issues in urban areas all over the world.Although the five most congested cities are located in developing countries, recent data indicatethat they are followed by Rome, Paris and London, which all present higher values than somelarge US cities1 (INRIX, 2019). In the case of Rome, for example, 166 hours are lost per driverdue to congestion per year. Evidence is growing that traffic congestion has many negativeconsequences related to employment (Hymel, 2009), pollution and health (Green et al., 2020,Simeonova et al., 2020, Requia et al., 2018) and road fatalities (Pasidis, 2019). All in all, thecurrent cost of road congestion in Europe is estimated to be over 270 billion euro per year (about1% of its GDP).

Several different options exist for addressing congestion in cities. One of the most commonpolicies has been to expand highway capacity. At times when politicians intend to foster economicgrowth, the increase in investment in road transportation has an important impact as a counter-cyclical fiscal policy (Leduc and Wilson, 2017). However, one of the main criticisms levelled at theexpansion of an intrametropolitan highway network is that this policy may not generate any realimprovements in accessibility and in economic growth; the evidence shows that, in the long term,these investments may simply relocate economic activity and leave congestion levels unchanged(Duranton et al., 2020).

One of the reasons for the inability of these policies to reduce urban congestion is the induceddemand effect, also known as the ‘fundamental law of highways congestion’ (Downs, 1962, 1992).As a result of the new demand induced by the added road capacity, the travel speed (as a measureof congestion) on an expanded highway reverts to its level prior to its expansion. The induceddemand phenomenon has undergone extensive empirical testing; however, the evidence is notconclusive for either short-run or long-run analyses (for an overview, see Hymel, 2019). Manyof the earlier studies lack a good identification strategy for explaining the causal links betweenthe increase in highway capacity and its impact on vehicle traffic. Interesting exceptions are thepapers by Duranton and Turner (2011) and Hsu and Zhang (2014), which show an elasticity oftraffic with respect to highway lane miles of approximately one for both US and Japanese urbanareas. A unit elasticity suggests that increasing capacity supply does not reduce traffic congestion,not even partially.

It is not clear that the same induced demand effect also holds for the cities in Europe. Thesecities are more compact than most US cities, and they are also characterized by a lower degree ofcar-dependency and the widespread use of public transportation2. While EU Regional and Co-hesion Funds have funded a considerable portion of the immense highway network development

1Taking into account the commute delay attributable to congestion delay, the 10 most congested cities in the worldin 2019 were Bogotá, Rio de Janeiro, Mexico City, Istanbul, Sao Paulo, Rome, Paris, London, Boston and Chicago(INRIX, 2019).

2According to OECD data, the average urban population density in the European metropolitan areas in 2011 was718 persons per km2, compared to only 282 in the US. OECD and Eurostat statistics indicate for the same year that caruse in Europe was some 42% lower than in the US. Also Europe is the world’s leader in public transit systems: twothirds of the large European cities have subways compared with only a third in the US (Gonzalez-Navarro and Turner,2018).

1

in the last few decades, to our knowledge, there are no integrated studies analysing the impactof highway expansion on traffic congestion in the cities of Europe as a whole.

In this paper we test the ’fundamental law of highway congestion’ for 545 metropolitan areasof the EU28 countries during the period 1985—2005. This analysis for the whole of Europe ismethodologically challenging because we need to overcome several identification issues in orderto properly estimate the causal effect. The first contribution of this paper is that we combineGIS data for a variety of historical data of transportation networks in Europe in order to obtainunbiased estimates. A second contribution is that we break down the effect of the highwayexpansion to the capacity effect (number of lanes) and the coverage effect (length of the network).Finally, our third contribution is that we are able to study the heterogeneity of the effects basedon the existence of other policies that target congestion in European cities. To see whether thesepolicies affect congestion, we first control for the road pricing policies, and then take into accountthe availability of public transportation (railroads and subways) connecting some city centers withtheir suburbs. Our average results indicate that, in line with the evidence from the US and Japan,for the European cities the induced demand effect is clear for both the capacity and coverageexpansion. However, we found that the induced demand is considerable smaller in cities withroad pricing schemes, and that congestion falls with public transportation expansions.

The remainder of the paper is organized as follows. In Section 2, we begin by describing ourdata on congestion and highways in European cities and how we process them. Then in Section3, we discuss our estimation approach and in Section 4 we present the results. To take account ofsome of the specific features of European cities, in Section 5 we include congestion pricing andthe availability of public transportation in European cities in the analysis. Finally, we present themain conclusions in Section 6.

2. Congestion and highways in Europe

We use the Functional Urban Area (FUA) (formerly known as Larger Urban Zone (LUZ)), definedby the European Commission (Urban Audit Project) and the OECD as the unit of observation. Incommon with the Metropolitan Statistical Area in the US, the FUA consists of a central city (withat least 50,000 inhabitants) and a commuting zone (made up of all municipalities with at least 15%of their employed residents working in the city. The final dataset includes 545 FUAs covering thewhole geography of Europe (29 countries)3.

2.1 Congestion in Europe

We use data from the Road Traffic Censuses conducted by the United Nations Economic Com-mission for Europe (UNECE). These censuses contain traffic and inventory information on themain highways in Europe (E-Road network) at a very detailed geographical level (road segments)for every five years from 1985 to 20054. Specifically, we obtain information on the annual average

334 cities have not been considered in the final sample due to the lack of congestion data.4Unfortunately, more recent censuses for 2010 and 2015 do not cover all European countries, only some of them.

2

daily traffic (AADT), the length (km) and the number of lanes of each segment for the years 1985,1995 and 2005.

To measure traffic congestion, we use the well-known indicator ’vehicle kilometers traveled’(VKT), that is, the kilometers traveled by motor vehicles on the highway network. We firstcompute its value at the segment level by multiplying the length of each highway segment (km)with its AADT. Then, we compute the VKT at the FUA level by summing the values for allhighway segments located within each FUA.

Table 1 reports the computations of these indicators for our sample of 545 European cities:Average values for both AADT and VKT, and individual VKT values for the top 10 and the bottom10 cities with population over 1 million inhabitants. Regarding the AADT, an average Europeancity in 2005 had 15,900 vehicles passing any point of the highway network. Between 1985 and2005, the number of vehicles increased by 70%. The high standard deviations in parenthesisindicate that the FUAs are quite different and heterogeneous in terms of AADT. Figures 1a and1b show that the increase of the AADT was a general feature in all Europe.

Table 1: Highway congestion in European cities, 1985–2005

1985 1995 2005 1985–1995 1995–2005 1985–2005

Annual average daily traffic (vehicles) - FUA 15,900 21,358 27,020 34.3% 26.5% 69.9%(15,318) (19,237) (21,837)

Vehicle kilometers traveled (’000 km) - FUA 2,673 3,603 4,586 34.8% 27.2% 71.6%(3,959) (4,722) (5,426)

VKT for the top 10 FUAs with population over 1 million (’000 km)London (UK) 57,435 62,633 66,830 9.1% 6.7% 16.4%Madrid (ES) 21,754 20,737 39,644 -4.7% 91.2% 82.2%Ruhrgebiet (DE) 20,326 28,405 29,964 39.7% 5.5% 47.4%Frankfurt (DE) 24,217 27,500 28,810 13.6% 4.8% 19.0%West Midlands (UK) 25,282 25,578 28,608 1.2% 11.8% 13.2%Berlin (DE) 13,610 20,425 24,930 50.1% 22.0% 83.1%München (DE) 18,386 19,747 23,111 7.4% 17.0% 25.7%Amsterdam (NL) 10,969 22,204 22,304 102% 0.5% 103 %Hamburg (DE) 17,212 22,548 22,169 31.0% -1.7% 28.8%Manchester (UK) 16,368 19,421 21,231 18.7% 9.3% 29.7%

VKT for the bottom 10 FUAs population over 1 million (’000 km)Leipzig (DE) 4,632 5,849 6.954 26.3% 18.9% 50.1%Napoli (IT) 3,526 5,877 5,798 66.7% -1.3% 64.5%Torino (IT) 1,774 5,272 5,289 197% 0.3% 198%Porto (PT) 1,081 3,443 5,125 218% 48.8% 374%Sofia (BG) 2,745 3,495 5,039 27.3% 44.2% 83.6%Ostrava (CZ) 902 1,227 3,825 36.0% 212% 324%Krakow (PL) 1,732 2,127 3,753 22.8% 76.5% 117%Katowice (PL) 1,012 1,856 3,708 83.5% 99.8% 267%Gdansk (PL) 1,419 1,591 2,806 12.1% 76.4% 97.7%Bucuresti (RO) 738 1,051 2,135 42.5% 103% 189%

Notes: FUA’s values are averages. Standard deviations in parenthesis.

VKT figures in Table 1 and Figures 1c and 1d also show high levels of traffic on highwaysand a growth between 1985 and 2005. On average, motor vehicles traveled 2.7 and 4.6 millionkm in 1985 and 2005, respectively. The standard deviations in parenthesis indicate that the trafficlevels were quite different between cities. Focusing on the most populated cities, the top 10 cities

3

had VKT values above the 20 million km. Their traffic grew with rates ranging from 13% inWest Midlands (UK) to 103% in Amsterdam (NL). The cities with lowest VKT had values in 2005between 2 million km (Bucuresti, RO) and almost 7 million km (Leipzig, DE). All these citiesexperienced significant increases in their traffic levels, between 50% (Leipzig, DE) and more than300% (Ostrava, CZ; Porto, PT).

Figure 1: Highway congestion in Europe

(a) 1985 AADT (b) 2005 AADT

(c) 1985 VKT (d) 2005 VKT

Table 2: Highways in European cities, 1985–2005

1985 1995 2005 1985–1995 1995–2005 1985–2005

Total length (km) - Europe 28,196 52,077 74,219 84.7% 74.2% 163%

Total length (km) - FUA 14,747 23,070 35,328 56.4% 53.1% 140%

Average length (km) - FUA 127 139 184 9.4% 32.3% 44.9%(106) (107) (146)

Average number of lanes - FUA 1.6 1.9 2.1 18.7% 10.5% 31%(0.9) (1.0) (1.2)

Average Lane kilometers - FUA 1,666 1,762 1,884 5.8% 6.9% 13.1%(1,441) (1,501) (1,575)

Notes: Standard deviations in parenthesis.

4

Figure 2: Highways in Europe

(a) Evolution of the highway network, 1985-2005 (b) 2005 Highway length

(c) 2005 Highway lanes (d) 2005 Highway lane kilometers

(e) 2005 Highway tolls

2.2 Highways in Europe

To measure the size of the highway network, we compute the so-called ’lane kilometers’. First, wemultiply the length (km) of each highway segment with the number of lanes. Then, we sum theresulting values for segments located within each city. As above mentioned, the UNECE RoadTraffic Censuses also provide information on segment length and lanes.

5

Table 2 presents the main characteristics of the highway network in 1985, 1995 and 2005 inEurope and in our sample of 545 cities. Of the more than 74,000 km of European highways in2005, almost half of them were located in FUAs. Between 1985 and 2005, the network more thandoubled. Figure 2a shows the evolution of the highway network in Europe between 1985 and2005.

Table 2 also reports average computations for the 545 FUAs and shows that highways wereextended both in terms of their coverage and their capacity. First, the length of the highwaynetwork of an average FUA increased from 127 to 184 km (a 45%) between 1985 and 2005. Second,the average number of lanes also increased from 1.6 to 2.1 in 20 years. As a result, the number oflane kilometers of the average FUA increased from 1,666 in 1985 to 1,884 km in 2005. The highstandard deviations of these three variables (in parenthesis) and their related maps in Figures 2b,2c and 2d indicate that the European cities show a high degree of heterogeneity.

Figure 2e shows a categorization of the share of tolled highway kilometers in the 545 FUAs. In20055, an average city had a 25% of tolled highways. However, our sample includes cities withouttolls (285) and with tolls (260). Among the latter, 77 cities have tolls in all their highways.

According to Albalate and Bel (2012), tolls in Europe are mostly used to finance constructionand maintenance costs of highways and, as a result, they are not related to traffic congestion.However, more recently some cities have adopted congestion pricing policies -the so-called urbantolls- in order to reduce the impact of traffic (congestion, noise and pollution). Figure 2e shows the14 cities that, according to https://www.urbanaccessregulations.eu, apply ’congestion prices’in 2020.

3. Empirical strategy

3.1 Pool, panel and first-difference

In this section, we introduce the empirical framework used to estimate the effect of the highwaynetwork expansion on the level of congestion. Increasing the supply of highways is expectedto lower the cost of motor vehicle use in the short run because of the increase in the overallhighway capacity in a city, which decreases traffic congestion. However, this reduction in themajor component of the cost of motor vehicle use might affect the travel decisions of individualsregarding the mode and quantity of travel. The ’fundamental law of highway congestion’ suggeststhat the long term average effect of increasing the supply of roads will be that induced demandwill bring the level of congestion back to its initial level (or even to a higher level).

To empirically study the role played by highways improvements on highway congestion, weindex cities by i and years by t, and estimate the following equation:

ln(VKTit) = β0 + β1 × ln(Lane kmit)

+ β2 × ln(Populationit) + β3 × Socioeconomyit+ β4 ×Geographyi + β5 ×Historyi + eit

(1)

5Unfortunately, only the 2005 UNECE Road Traffic Census provide this kind of information at the highway segmentlevel.

6

https://www.urbanaccessregulations.eu

where VKTit refers to the vehicle kilometers traveled and measures the kilometers traveled bymotor vehicles in the highway network. The main explanatory variable is Lane kmit and refers tothe number of highway lane kilometers. Populationit is the number of inhabitants from officialpopulation censuses provided by the DG REGIO of the European Commission. Socioeconomyit isa vector of characteristics including income, proxied by GDP; unemployment rate; and industrialcomposition, proxied by the shares of employment in manufacturing, in financial and businessservices, and in non-market services. Since there are no data available at the FUA level, all threevariables are computed using data from the NUTS3 in which the FUA is located. Geographyiincludes controls for physical geography such as total land area (km2); a suburbanization index,which is the share of the central city area; the altitude (meters), elevation range (meters), and theterrain ruggedness index computed a la Riley et al. (1999) using the Digital Elevation Model overEurope (EU-DEM)6; and the logarithm of the distance (m) to the closest coast from the centroid ofthe central city. Finally, Historyi adds two types of historical controls. First, dummy variables forhistorical major cities in 814, 1000, 1200, 1450 and 18507. Second, three controls for more recenthistory (the 20th century): The logarithms of the decennial levels of population between 1960 and1980.

With data describing a panel of cities, we can partition eit into permanent (δi) and time-varying(ηit) components. By so doing, we can remove all time-invariant city effects by estimatingEquation (1) using city fixed-effects (Equation (2)) or its first-difference version (Equation (3)):

ln(VKTit) = β1 × ln(Lane kmit)

+ β2 × ln(Populationit) + β3 × Socioeconomyit + δi + ηit(2)

∆ln(VKTit) = β1 × ∆ln(Lane kmit)

+ β2 × ∆ln(Populationit) + β3 × ∆Socioeconomyit + ∆ηit(3)

where ∆ is the first-difference operator.

3.2 Instrumental variables

When the random element of traffic congestion is uncorrelated with highways, we can estimateEquations (1), (2) and (3) by ordinary least squares (OLS). However, highway lane kilometers areexpected to be endogenous to vehicle kilometers traveled because of reverse causation (e.g., con-gestion fostering highway expansion), measurement error (e.g., highways mismeasured becausesome may have just opened or are about to be opened) and omitted variables (e.g., geography,amenities or economic structure leading to more highways).

To address concerns of endogeneity, we rely on IV estimations (limited-information maximumlikelihood, LIML). We use the digital vector maps created or used by Garcia-López (2019) tobuilt instruments based on the ancient road and railroads in Europe: The main and secondaryroads during the Roman Empire (McCormick et al., 2013), the main trade routes in the Holy

6The original GIS raster maps are available in https://www.eea.europa.eu/data-and-maps/data/copernicus-land-monitoring-service-eu-dem.

7These dummies are computed using information from the Digital Atlas of Roman and Medieval Civilizations(DARMC, http://darmc.harvard.edu) and from Bairoch et al. (1988).

7

https://www.eea.europa.eu/data-and-maps/data/copernicus-land-monitoring-service-eu-demhttps://www.eea.europa.eu/data-and-maps/data/copernicus-land-monitoring-service-eu-demhttp: //darmc.harvard.edu

Roman Empire in the 15th century (Ciolek, 2005), the postal roads in 1810 according to A.Arrowsmith’s map8, and the railroad network in 1870 (Marti-Henneberg, 2013). Figures A.1and A.2 in Appendix A.1 shows their location in Europe and Barcelona, respectively.

Since, by definition, these historical instruments are time-invariant, we follow Baum-Snow(2007) and Garcia-López (2012, 2019) and adopt a ’shift-share’ approach a la Bartik (1991) usingeach historical (rail)road as the ’share’ component and the evolution of the highway networkas the ’shift’ component. Specifically, we compute each time-variant historical instrument bymultiplying its historical length (in km) by the fraction of the highway network kilometragein each country completed at each year and excluding each city’s own contribution.

As common sense suggests, historical transportation networks may be relevant because mod-ern networks are not built in isolation from them. On the contrary, it is easier and cheaper to buildnew infrastructures close to old infrastructures. Duranton and Turner (2011, 2012), Garcia-López(2012) or Garcia-López et al. (2015), among others, show that the stocks of historical (rail)roads areindeed highly correlated with the stocks of modern transportation networks in the US, Barcelonaand Spain, respectively.

We econometrically test the relevance of each time-variant historical (rail)road in AppendixA.2. Specifically, Table A.1 shows OLS results when we regress the highway lane kilometers onthe length of each ancient (rail)road (km) following a pooled strategy (columns 1 to 3, Equation(A.1)) and panel fixed-effect strategy (columns 4 and 5, Equation (A.2)). We also follow a first-difference strategy regressing the changes in the highway lane kilometers on changes in the lengthof each ancient (rail)road (km) (columns 6 and 7, Equation (A.3)), and adding the lag of VKTwhile controlling for geography, history and country fixed-effects (columns 8 and 9) or for FUAfixed-effects (columns 10 and 11). According to Murray (2006), valid instruments should havesignificant effects on modern highway lane kilometers and high First-Stage F-statistics. First-stageresults show that both Roman roads and the 1870 railroads predict the congestion level when wefollow a pooled regression strategy (columns 1 and 2). However, when we use a panel fixed-effectand first-difference regressions (columns 4, 6, 8 and 10), only the Roman roads predict both thelevels and the changes in the highway lane kilometers. Results also show First-Stage F-statisticsare near or above Stock and Yogo (2005)’s critical values, in particular when only Roman roadsare used as instrument.

Our time-variant historical instrument also needs to be exogenous. As Garcia-López (2019)explains in detail, its ’shift’ element is exogenous because, by construction, it refers to the lengthof the highway network that would have existed in each year had governments allocated highwayconstruction uniformly across Roman roads within the countries. The ’share’ component, thelength of Roman roads, may also be exogenous because Roman roads were not built to anticipatethe current traffic congestion levels in European cities. On the contrary, they were built to achievemilitary, administrative, and commercial goals between different parts of the Roman Empire(Garcia-López et al., 2015, Garcia-López, 2019). However, since geography and history haveinfluenced both the evolution of European cities and its transportation networks, the exogeneityof the Roman instrument hinges on controlling for those characteristics as in the pooled regression

8See the David Rumsey Map Collection (http://www.davidrumsey.com) for the original paper maps.

8

http://www.davidrumsey.com

strategy (Equation (1)). Alternatively, we can add city fixed-effects as in Equation (2) or estimatea first-difference regression as in Equation (3).

4. Does the fundamental law of highway congestion apply in Europe?

4.1 Main results

The fundamental law of highway congestion has been confirmed in the US (Duranton and Turner,2011) and Japan (Hsu and Zhang, 2014). In this section, we investigate whether this (more than)proportional increase in congestion levels when highways are expanded also applies in Europe.

Table 3 reports OLS and LIML results in columns 1 to 5 and 6 to 7, respectively, when weregress the log of vehicle kilometers traveled on the log of highway lane kilometers. In columns1 and 6, we follow a pooled strategy and estimate Equation (1) without control variables. Incolumns 2 and 7, we add all control variables. Columns 3 to 5 and 8 to 10 show results when wefollow a fixed panel strategy and estimate Equation (2) gradually adding time-variant explanatoryvariables: the log of lane kilometers in columns 3 and 8, the log of population in columns 4 and9, and socioeconomic controls in columns 5 and 10. Table 3 also reports first-stage statistics forthe LIML regressions (which use the log of the time-variant length of Roman roads as instrument),and all of them are near or above Stock and Yogo (2005)’s critical values.

Table 3: The effect of highways on traffic congestion, OLS and IV results

Dependent variable: ln(VKT)

OLS results IV results

Pool Panel Pool Panel

[1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

ln(Lane km) 0.969a 0.831a 0.689a 0.686a 0.695a 1.480a 1.558a 1.256a 1.285a 1.206a

(0.026) (0.034) (0.096) (0.097) (0.097) (0.150) (0.272) (0.304) (0.337) (0.314)ln(Population) X X X X X XGeography X XHistory X XSocioeconomy X X X XCountry fixed-effects X XFUA fixed-effects X X X X X XYear fixed-effects X X X X X X X X X X

Adjusted R2 0.80 0.89 0.70 0.70 0.71First-Stage F-statistic 29.51 12.01 21.12 19.90 19.79Instrument ln(Km of Roman r.) ln(Km of Roman roads)

Notes: 1,635 observations (545 cities × 3 decades (1985-2005)) in each regression. Geography controls include the logarithm of theFUA area, a suburbanization index, which is the share of the central city area, the mean and range of FUA elevation, the meansurface ruggedness for each FUA and the logarithm of the distance to the closest coast from the central city centroid. Historyincludes the logarithms of FUA population in 1960, 1970 and 1980, and dummy variables for historical major cities in 814, 1000,1200, 1450 and 1850. Socioeconomic characteristics are the log of the GDP, the share of employment in manufacturing, the shareof employment in finance and business services, the share of employment in non-market services, and the unemployment rate.Robust standard errors are clustered by FUA and are in parenthesis. a, b and c indicates significant at 1, 5, and 10 percent level,respectively.

9

The estimated OLS coefficient of interest is positive and significantly different from zero inall specifications and shows that a 1% increase in the log of lane kilometers increases the vehiclekilometers traveled between 0.7 (column 5) and 0.8% (column 2). After addressing concernsof endogeneity, LIML counterparts in columns 7 to 10 show a higher impact of highways oncongestion between 1.2 and 1.6%, respectively.

Appendix B provides additional results following a pooled strategy. In Table B.1, we estimateEquation (1) for each year: 1985, 1995 and 2005. OLS and LIML results are between 0.8-0.9 and1.3-1.8, respectively, and are not statistically different from their counterparts in Table 3. In TableB.2, we show that the estimated coefficient of interest is quite stable when we gradually addexplanatory variables to the pooled regression.

We also perform a set of robustness checks in Table C.1 of Appendix C. First, to addressendogeneity concerns about the population variable (Duranton and Turner, 2011), in Panel A wefollow the pooled strategy and estimate Equation (1) instrumenting the population variable withtwo time-invariant instruments: the average temperature and the average precipitation of theFUA. In column 1 we only instrument population, in column 2 we simultaneously instrument thetwo endogenous variables. In both cases, the the associated First-Stage statistics are above andclose to the Stock and Yogo (2005)’s critical values, respectively. Furthermore, the OLS and LIMLresults are in line with those in columns 2 and 7 of Table 3.

Second, we are worried that some country specific shocks may have affected both the evolutionof highways and congestion at the country level. To allow the countries to have different timetrends, in Panel B we follow the panel fixed-effect strategy and estimate Equation (2) addingcountry-specific linear trends in columns 3 and 4. Both OLS and LIML results hold.

Third, another potential concern is that the increase in highway traffic and the highwaydevelopment are both affected by the supply of other roads that are not classified as highways9.In Panel C, we estimate Equation (2) adding the length of secondary roads and the length oflocal roads as explanatory variables. Since these two additional variables are also endogenous,we instrument them in column 6 using the time-variant lengths of the 15th c. trade routes duringthe Holy Roman Empire and of the 1810 post routes10. The related First-Stage F-statistic is nearthe Stock and Yogo (2005)’s critical value. Results for the lane kilometers are not statisticallydifferent from those in columns 2 and 7 of Table 3. Interestingly, the expansion of secondaryroads decreases highway congestion, but the effect is quite small (0.02-0.03).

Finally, we also test for the functional form of the effect under study. One might expect that theeffect of highway expansion on traffic congestion depends crucially on the extent of the highwaynetwork in each city. In Panel D, we add the square of the log of lane kilometers. Resultsshow that this quadratic term is not statistically significant and its value is very close to zero.Furthermore, the estimated coefficient for lane kilometers is in line with previous results.

Now we change our empirical strategy and follow the first-difference approach by estimatingEquation (3). Table 4 presents OLS and LIML results in columns 1 to 5 and 6 to 10, respectively.

9We use data of secondary and local roads provided by the EC DG-REGIO (for more details, see Stelder, 2016).10We build these instruments using the ’shift-share’ approach explained in Section 3.2. The length of each historical

(rail)road is the ’share’ component and the evolution of the country transportation network (excluding each city’s owncontribution) is the ’shift’ component.

10

Columns 1 to 3 and 6 to 8 show results when we gradually add time-variant explanatory variables:Only the change in lane kilometers (columns 1 and 6), adding the change in population (columns2 and 7) and including the changes in socioeconomic controls (columns 3 and 8). Since all time-invariant factors drop out of Equation (3), we add geography, history and country dummies ascontrol variables in columns 4 and 9, and add FUA fixed-effects in columns 5 and 10. Moreover,we also add the logarithm of the lagged vehicle kilometers traveled in columns 4-5 and 9-10. Table4 also reports first-stage statistics for the LIML regressions (which use the log of the time-variantchange in the length of Roman roads as instrument), and all of them are above Stock and Yogo(2005)’s critical values.

Results for the OLS regressions show elasticities of the change in lane kilometers rangingbetween 0.2 in the most demanding specification (column 5) and 0.6 in the pure first-differencespecification (column 3). After instrumenting lane kilometers with the time-variant Roman roadinstrument, LIML results for the different specifications are quite stable and show an elasticityaround 1.2%. Both OLS and LIML results are in line with those using pooled and panel fixed-effect approaches in Table 3.

Table 4: The effect of highways on traffic congestion, OLS and IV results: First-difference

Dependent variable: ∆ln(VKT)

OLS results IV results

[1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

∆ln(Lane km) 0.632a 0.633a 0.638a 0.520a 0.243a 1.190a 1.177a 1.249a 1.104b 1.138a

(0.103) (0.103) (0.103) (0.092) (0.063) (0.328) (0.361) (0.375) (0.438) (0.306)

Lagged ln(VKT) -0.161a -1.057a -0.131a -0.874a

(0.020) (0.048) (0.039) (0.088)∆ln(Population) X X X X X X X XGeography X XHistory X X∆Socioeconomy X X X X X XCountry fixed-effects X XFUA fixed-effects X X

Adjusted R2 0.32 0.32 0.33 0.41 0.80First-Stage F-statistic 23.58 22.01 21.26 14.12 16.22Instrument ∆ln(Km of Roman roads)

Notes: 1,090 observations (545 cities × 2 first-difference periods) in each regression. Geography controls include the logarithm ofthe FUA area, a suburbanization index, which is the share of the central city area, the mean and range of FUA elevation, the meansurface ruggedness for each FUA and the logarithm of the distance to the closest coast from the central city centroid. Historyincludes the logarithms of FUA population in 1960, 1970 and 1980, and dummy variables for historical major cities in 814, 1000,1200, 1450 and 1850. Socioeconomic characteristics are the log of the GDP, the share of employment in manufacturing, the shareof employment in finance and business services, the share of employment in non-market services, and the unemployment rate.Robust standard errors are clustered by FUA and are in parenthesis. a, b and c indicates significant at 1, 5, and 10 percent level,respectively.

In summary, the fundamental law of highway congestion states that a capacity expansionof highways produces a (more than) proportional increase in highway congestion levels. LIMLresults when using pooled, panel and first-difference empirical strategies confirm that this lawalso applies in European cities. Focusing in our preferred specification in column 10 of Table 3

11

that follows a panel fixed-effect estimation, a 1% increase in the highway lane kilometers increasesthe vehicle kilometers traveled by 1.2%. This elasticity is slightly higher than the one obtained byDuranton and Turner (2011) for the US (1.02) and almost the same than the estimated by Hsu andZhang (2014) for Japan (1.24).

4.2 Coverage and capacity effects

A highway network can be extended by increasing the number of lanes, that is, its ’capacity’.Another way is extending the length of existing routes or creating new ones, that is, increasingthe ’coverage’ of the highway network. After confirming that the fundamental law apply toEuropean cities, we now turn our attention to study whether the effect of an increase in highwayprovision is related to a ’coverage effect’ and/or to a ’capacity effect’.

Based on previous main results, we depart from our preferred panel fixed-effect approachto estimate an equation that consider both the length of the highway network and the averagenumber of highway lanes11:

ln(VKTit) = β1a × ln(Km of highwaysit)

+ β1b ×Average number of lanesit+ β2 × ln(Populationit) + β3 × Socioeconomyit + δi + ηit

(4)

It is important to notice that these two variables are also endogenous and, as a result, needto be instrumented. For the case of the length of highways, Garcia-López (2019) shows thathistorical (rail)roads and, in particular, Roman roads, are good predictors of this variable. TableA.2 of Appendix A.3 reports first-stage OLS results when we regress the log of km of highwayson the log of km of our time-variant historical (rail)roads conditional on controls. In column 1,we separately include the four old (rail)roads and find that only Roman roads has a positive andsignificant effect. In column 2, we only include the Roman roads and confirm its effect. In bothcases, the First-Stage F-Statistic is near Stock and Yogo (2005)’s thresholds.

As shown by Garcia-López (2012), Garcia-López et al. (2015) and Garcia-López et al. (2017a,b),the location of modern highways is also related to the location of historical (rail)roads. In fact,when mapping these old (rail)roads we realize that most of them are located very close to eachother (see, for example, the case of Barcelona in Figure A.2). This could mean that, conditionalon the geography, building additional lanes is easier and cheaper in those highway segmentswhere we also find more old (rail)roads in parallel. Based on this idea, we instrument the averagenumber of highways lanes with the average number of historical (rail)roads12. Column 3 of TableA.2 shows first-stage results when we consider the average number of all historical (rail)roads. Incolumn 4, we separately consider the average number of each historical (rail)road. In both cases,estimated coefficients are positive and significant. The related First-Stage F-Statistics are abovethe Stock and Yogo (2005)’s thresholds.

11We consider the number of lanes in each highway segment and compute the average for each city.12For each year and highway segment, we compute the number of historical (rail)roads that are located in parallel

less than 1 kilometer. Then, we compute the average for each city.

12

Table 5 reports LIML results of estimating different specifications of Equation (4). First, weseparately consider the logarithm of the highway length (column 1) and the average number oflanes (columns 2 and 4). Later, we jointly include both variables in columns 3 and 5. Sinceboth variables are expected to be endogenous, we instrument them with the above mentionedinstruments: The length of historicals (rail)roads (columns 1, 3 and 5), the average number ofall historical (rail)roads (columns 2 and 3), and the average number of each historical (rail)road(columns 4 and 5). These LIML results show positive effects of extending highway routes andincreasing lanes. In particular, our preferred results in column 3 indicate that doubling thehighway network, either by increasing the highway length by 100% or by building one additionallane in all highway segments, causes a 114% and a 212% growth in the vehicle kilometers traveled,respectively. These elasticities confirm both the coverage and capacity effects, being the lattermore important.

Table 5: The effect of highways on traffic congestion, IV results: Length and lanes


Instrumenting: Length Lanes Both Lanes Both[1] [2] [3] [4] [5]

ln(Km of highways) 2.293a 1.146a 1.287a

(0.561) (0.288) (0.208)

Number of lanes 2.117a 2.198a 2.702a 2.540a

(0.493) (0.369) (0.513) (0.333)ln(Population) X X X X XSocioeconomy X X X X XFUA fixed-effects X X X X XYear fixed-effects X X X X X

First-Stage F-statistic 11.52 26.29 7.65 11.78 8.47Instruments ln(Roman roads) ln(Roman roads) ln(Roman roads)

Number of all historical (rail)roads Numbers of each historical (rail)road

Notes: 1,635 observations (545 cities × 3 decades (1985-2005)) in each regression. Socioeconomic characteristics are the log ofthe GDP, the share of employment in manufacturing, the share of employment in finance and business services, the share ofemployment in non-market services, and the unemployment rate. Robust standard errors are clustered by FUA and are inparenthesis. a, b and c indicates significant at 1, 5, and 10 percent level, respectively.

5. Do tolls and public transportation matter?

Compared with US cities, there are at least two prominent features in European cities: Theuse of tolls, and the massive investment in public transportation, in particular, in railroads. Inthis section, we study whether these two characteristics affect the fundamental law of highwaycongestion.

5.1 Tolls

As discussed in Section 2.2, in 2005 almost half of the cities included in our sample had tolls(260 out of 545) and, on average, these tolls affected 25% of the highway network. To analyze therole of these tolls, we depart from the panel fixed-effect approach to estimate an equation that

13

includes the interaction of our main explanatory variable:


+ β2 × ln(Lane kmit)× Tollsi+ β3 × ln(Populationit) + β4 × Socioeconomyit + δi + ηit

(5)

Table 6 show results for Equation (5). In column 1, we interact lane kilometers with a dummyfor cities with tolled highways. In column 2, the interaction variable is a dummy for cities with25% or more of tolled highways. Finally, we directly interact lane kilometers with the shareof tolled highway segments in columns 3. Results are essentially identical and show that (1)highway improvements increase congestion, (2) the effect is smaller in cities with tolls, and(3) the fundamental law is mainly related to cities without tolls or with a low percentage oftolled highways. In particular, and focusing on our preferred specification in column 3 (using acontinuous interaction), a 1% increase in lane kilometers increases congestion by 1.9% in citieswithout tolls and by only 0.3% (=1.9-1.6) in cities with tolls in all their highways (100% shareof tolled highways). Some simple computations show that the fundamental law applies to citieswith a share of tolled highways below 56%. These results can be regarded as novel evidencein line with recent literature suggesting that the solution to traffic congestion is the adoption of’congestion’ pricing policies (Santos, 2004, de Palma et al., 2006, Winston and Langer, 2006, Leape,2006, Eliasson and Mattsson, 2006).

Table 6: The effect of highways on traffic congestion, IV results: Tolls and interactions

Dependent var.: ln(VKT)

Interacting: All tolled cities Above 25% tolls cities Share of tolled highways

[1] [2] [3]

ln(Lane km) 1.903a 2.106a ln(Lane km) 1.894a

(0.560) (0.743) (0.629)

ln(Lane km) -1.203b -1.454b ln(Lane km) -1.598b

×Dummy tolls (0.527) (0.685) ×Share tolls (0.785)ln(Population) X X ln(Population) XSocioeconomy X X Socioeconomy XFUA fixed-effects X X FUA fixed-effects XYear fixed-effects X X Year fixed-effects X

First-Stage F-stat 11.18 7.62 First-Stage F-stat 4.50Instruments ln(Km of Roman roads) ln(Km of Roman roads) Instruments ln(Km of Roman roads)

ln(Roman)×Tolls ln(Roman)×25% Tolls ln(Roman)×Share tolls


We might fear that the allocation of highway tolls is endogenous to traffic. However, asdiscussed in Section 2.2, tolls in Europe are mostly used to finance construction and maintenancecosts of highways (Albalate and Bel, 2012). Only recently some cities have adopted ’congestion’pricing policies in order to reduce the impact of traffic (congestion, noise, pollution, ...). According

14

to https://www.urbanaccessregulations.eu, only 14 of our 545 cities apply ’congestion prices’in 2020. The above results hold when we do not include the observations related to these 14 cities.

Finally, in Appendix D we follow an alternative empirical strategy based on estimating Equa-tion (2) for two subsamples: Cities with tolls vs. without tolls. Results reported in Table D.1confirm that highway tolls mitigate the induced demand effect of highway expansions.

5.2 Public transportation

Now we turn our attention to public transportation and, in particular, the railroad network. InEurope, it was mainly built during the 19th and 20th centuries and nowadays there are morethan 225,000 km of rail lines. At the FUA level, the network expanded from 32,000 to 52,000 kmbetween 1985 and 2005. On average, the network increased more than 60%, from 59 to 96 km.Some examples of the more recent railroad expansion are the construction of the Regional ExpressRail (RER) network in Paris (Garcia-López et al., 2017a,b) or the introduction of the high-speedrail in Germany (Ahlfeldt and Feddersen, 2018), Spain (Albalate et al., 2017) and the UK (Heblichand Simpson, 2018).

We study the effect of railroads on congestion by estimating a version of Equation (2) whichincludes the length of railroads as explanatory variable:


+ β2 × ln(Km of railroadsit)

+ β3 × ln(Populationit) + β4 × Socioeconomyit + δi + ηit

(6)

Since the stock of railroads may also depend on the level of highway congestion, now we haveto deal with an additional endogenous variable. To do so, we follow Garcia-López (2019), whichshows that the 1870 railroad network is a good predictor of the length of modern railroads. TableA.3 of Appendix A.4 reports first-stage OLS results when we regress the log of km of railroadson the log of km of our time-invariant historical (rail)roads conditional on controls. In column1, we separately include the four old (rail)road networks and find that only the 1870 railroadshas a positive and significant effect. In column 2, we only include the 1870 railroad networkand confirm its effect. In both cases, the First-Stage F-Statistic is near Stock and Yogo (2005)’sthresholds.

Column 1 of Table 7 reports LIML results when we estimate Equation (6). They confirm thepositive effect of lane kilometers on congestion and the fundamental law of highway congestionwith a VKT elasticity above 1. More interesting are the results for railroads: Their estimated coef-ficient is significant and negative, indicating that railroads directly moderate highway congestion.In particular, a 1% increase in the length of the railroad network decreases congestion by 0.5%.

Subways are important in the largest European cities because a high proportion of theirmobility is based on this transportation mode (Gonzalez-Navarro and Turner, 2018). In oursample, 32 cities have a subway network within their cores and connecting with their suburbs.The subway network of the average city had 52 km in 2011 and it represented around 14% of thetotal railroad length. To study whether subways also matter for congestion, we estimate Equation

15

https://www.urbanaccessregulations.eu

(6) including the interaction13 between the railroad length and the share of subways. Column2 shows LIML results when we only instrument lane km and railroad length. Our preferredspecification in column 3 reports LIML estimates when the interaction term is also instrumented.In this case, results confirm (1) the fundamental law of highway congestion (with a VKT elasticityabove 1) and (2) the above mentioned negative effect of railroads on congestion. Furthermore,the negative estimated coefficient for the interaction term shows that (3) the more important thesubways in the railroad network, the higher the reduction of congestion. In particular, a 1%increase in the length of the railroad network decreases congestion by 0.6% in a city withoutsubways, by 0.8% in a city with the average share of subways (14%), and by 1.3% in a city wheresubways are 50% of the total railroad network.

Table 7: The effect of highways on traffic congestion, IV results: Public transportation


Interacting with the share of subways

[1] [2] [3]

ln(Lane km) 2.208a 1.408a 1.407a

(0.531) (0.324) (0.344)

ln(Km of railroads) -0.533b -0.488b -0.562b

(0.270) (0.240) (0.268)

ln(Km of railroads) -0.660c -1.384c

×Share of subways (0.393) (0.760)ln(Population) X X XSocioeconomy X X XFUA fixed-effects X X XYear fixed-effects X X XShare of subways X X

First-Stage F-stat 4.98 3.31 2.12Instruments ln(Km of Roman roads) ln(Km of Roman roads) ln(Km of Roman roads)

ln(Km of 1870 railroads) ln(Km of 1870 railroads) ln(Km of 1870 railroads)ln(1870 Rail)×% Subways


In summary, the above results point out that railroads serve to free up highway capacity bytaking drivers off the highways and putting them in trains. These findings are in contrast to thoseof Duranton and Turner (2011), who find no effect of public transit on congestion. However, theirempirical evidence is based on regular buses, which are as prone to traffic congestion as motorvehicles. On the contrary, our findings are in line with those of Anderson (2014), who finds thatpublic transit generates large congestion relief benefits.

13We use an empirical strategy based on an interaction term (instead of directly using the length of the subwaynetwork as explanatory variable) because, unfortunately, we only have GIS maps and data for subways in 2011.

16

6. Conclusions

In this paper, we provide evidence that the ’fundamental law of highway congestion’ holds for thecities of Europe, but also that well designed public policies can moderate this congestion. We usedata for the 545 largest European cities to estimate the elasticity of a measure of congestion withrespect to highway expansion. The results indicate that this elasticity is in the range close to 1.This suggests that expansion of the highway network induced the demand for car travel, and so,on average, the level of congestion remained roughly unchanged in the period 1985–2005. In otherwords, we show that investments in highways did not effectively relieve traffic congestion. Wealso break down the type of expansion into coverage (the network length) and capacity (numberof lanes); we find that both effects were significant in explaining the increase in traffic especiallythe capacity effect. Controlling further for road pricing and public transportation policies, ourresults indicate that cities with these policies in place suffer less congestion.

These findings have major implications for transportation policies, given the persisting severityof traffic congestion in European urban areas where the reduction of car use has became a priority.Only a few cities have established congestion charges, and public transport systems are underfinancial pressure. It is clear that there is great deal of scope for the design of good policies.Related to our results, it is well known that public acceptance of road pricing schemes is verylow. People are not willing to pay in order to be able to drive and, on the other hand, theevidence seems to show that this policy might be regressive. A good way to proceed would be tointroduce urban tolls and use the revenues obtained to improve public transportation.

17

References

Ahlfeldt, G. and Feddersen, A. (2018). From periphery to core: Measuring agglomeration effectsusing high-speed rail. Journal of Economic Geography, 18(2):355–390.

Albalate, D. and Bel, G. (2012). Motorways, tolls and road safety: evidence from Europe. SERIEs,3(4):457–473.

Albalate, D., Campos, J., and Jiménez, J. L. (2017). Tourism and high speed rail in Spain: Doesthe AVE increase local visitors? Annals of Tourism Research, 65:71–82.

Anderson, M. L. (2014). Subways, Strikes, and Slowdowns: The Impacts of Public Transit onTraffic Congestion. The American Economic Review, 104(9):2763–2796.

Bairoch, P., Batou, J., and Chevre, P. (1988). Population of European cities from 800-1860. GenevaUniversity, centre of International Economic History, Geneva (CH).

Bartik, T. J. (1991). Who Benefits from State and Local Economic Development Policies? W.E. UpjohnInstitute for Employment Research.

Baum-Snow, N. (2007). Did highways cause suburbanization? The Quarterly Journal of Economics,122(2):775–805.

Ciolek, T. M. (2005). Georeferenced data set (Series 1 - Routes): Main trade routes in theHoly Roman Empire and nearby countries, c. 1500 ce. OWTRAD Dromographic DigitalData Archives (ODDDA). Old World Trade Routes (OWTRAD) Project. Canberra: Asia PacificResearch Online. www.ciolek.com/OWTRAD/DATA/tmcZCEm1500.html.

de Palma, A., Lindsey, R., and Proost, S. (2006). Research challenges in modelling urban roadpricing: An overview. Transport Policy, 13(2):97–105.

Downs, A. (1962). The Law of peak-hour Expressway Congestion. Traffic Quarterly, 16(3):393–409.

Downs, A. (1992). Stuck in Traffic: Coping with Peak-hour Traffic Congestion. Brookings InstitutionPress.

Duranton, G., Nagpal, G., and Turner, M. (2020). Transportation infrastructure in the US. WorkingPaper 27254, National Bureau of Economic Research.

Duranton, G. and Turner, M. A. (2011). The fundamental law of road congestion: Evidence fromUS cities. American Economic Review, 101(6):2616–2652.

Duranton, G. and Turner, M. A. (2012). Urban growth and transportation. Review of EconomicStudies, 79(4):1407–1440.

Eliasson, J. and Mattsson, L.-G. (2006). Equity effects of congestion pricing: Quantitative meth-odology and a case study for Stockholm. Transportation Research Part A: Policy and Practice,40(7):602–620.

Garcia-López, M.-A. (2012). Urban spatial structure, suburbanization and transportation inBarcelona. Journal of Urban Economics, 72(2–3):176–190.

Garcia-López, M.-A. (2019). All roads lead to Rome ... and to sprawl? evidence frome Eropeanciteis. Regional Science and Urban Economics, 79:103467.

18

www.ciolek.com/OWTRAD/DATA/tmcZCEm1500.html

Garcia-López, M.-A., Hémet, C., and Viladecans-Marsal, E. (2017a). How does transportationshape intrametropolitan growth? An answer from the Regional Express Rail. Journal of RegionalScience, 57(5):758–780.

Garcia-López, M.-A., Hémet, C., and Viladecans-Marsal, E. (2017b). Next train to the polycentriccity: The effect of railroads on subcenter formation. Regional Science and Urban Economics,67(C):50–63.

Garcia-López, M.-A., Holl, A., and Viladecans-Marsal, E. (2015). Suburbanization and highwaysin Spain when the Romans and the Bourbons still shape its cities. Journal of Urban Economics,85(C):52–67.

Gonzalez-Navarro, M. and Turner, M. A. (2018). Subways and urban growth: Evidence fromearth. Journal of Urban Economics, 108:85 – 106.

Green, C. P., Heywood, J. S., and Navarro Paniagua, M. (2020). Did the london congestion chargereduce pollution? Regional Science and Urban Economics, 84:103573.

Heblich, S. and Simpson, H. (2018). Commuting fast and commuting slow: The effects of highspeed rail. mimeo.

Hsu, W.-T. and Zhang, H. (2014). The fundamental law of highway congestion revisited: Evidencefrom national expressways in Japan. Journal of Urban Economics, 81:65–76.

Hymel, K. (2009). Does traffic congestion reduce employment growth? Journal of Urban Economics,65(2):127–135.

Hymel, K. (2019). If you build it, they will drive: Measuring induced demand for vehicle travelin urban areas. Transport Policy, 76(C):57–66.

INRIX (2019). Inrix global traffic scorecard. Technical report, INRIX Research.

Leape, J. (2006). The London Congestion Charge. The Journal of Economic Perspectives, 20(4):157–176.

Leduc, S. and Wilson, D. (2017). Are State Governments Roadblocks to Federal Stimulus? Evid-ence on the Flypaper Effect of Highway Grants in the 2009 Recovery Act. American EconomicJournal: Economic Policy, 9(2):253–292.

Marti-Henneberg, J. (2013). European integration and national models for railway networks(1840–2010). Journal of Transport Geography, 26(0):126–138.

McCormick, M., Huang, G., Zambotti, G., and Lavash, J. (2013). Roman road network (version2008). DARMC Scholarly Data Series, Data Contribution Series 2013-5. DARMC, Center forGeographic Analysis, Harvard University, Cambridge MA 02138. http://darmc.harvard.edu.

Murray, M. P. (2006). Avoiding invalid instruments and coping with weak instruments. Journal ofEconomic Perspectives, 20(4):111–132.

Pasidis, I. (2019). Congestion by accident? A two-way relationship for highways in England.Journal of Transport Geography, 76(C):301–314.

Requia, W. J., Higgins, C. D., Adams, M. D., Mohamed, M., and Koutrakis, P. (2018). The healthimpacts of weekday traffic: A health risk assessment of pm2.5 emissions during congestedperiods. Environment international, 111:164—176.

19

http://darmc.harvard.edu

Riley, S. J., DeGloria, S. D., and Elliot, R. (1999). A terrain ruggedness index that quantifies topographicheterogeneity, volume 5.

Santos, G. (2004). Road Pricing: Theory and Evidence. Elsevier.

Simeonova, E., Currie, J., Nilsson, P., and Walker, R. (2020). Congestion pricing, air pollution andchildren’s health. Journal of Human Resources, Forthcoming.

Stelder, D. (2016). Regional Accessibility Trends in Europe: Road Infrastructure, 1957-2012.Regional Studies, 50(6):983–995.

Stock, J. H. and Yogo, M. (2005). Testing for weak instruments in linear iv regression. In Andrews,D. W. and Stock, J. H., editors, Identification and Inference for Econometric Models: Essays in Honorof Thomas Rothenberg, pages 80–108. Cambridge: Cambridge University Press.

Winston, C. and Langer, A. (2006). The effect of government highway spending on road users’congestion costs. Journal of Urban Economics, 60(3):463–483.

20

Appendix A. Historical roads and railroads in Europe

A.1 Maps

Figure A.1: Historical roads and railroads in Europe

Figure A.2: Historical roads and railroads in Barcelona

21

A.2 First-stage results for highway lane kilometers

To econometrically test the relevance of each time-variant historical (rail)road, we estimate thefollowing first-stage equations and show their results in Table A.1:

ln(Lane kmit) = γ0 + γ1 × ln(Historical (rail)road kmit)

+ γ2 × ln(Populationit) + γ3 × Socioeconomyit+ γ4 ×Geographyi + γ5 ×Historyi + eit

(A.1)

when we follow a pooled regression approach (Equation (1)). Results are in columns 1 to 3.

ln(Lane kmit) = γ1 × ln(Historical (rail)road kmit)

+ γ2 × ln(Populationit) + γ3 × Socioeconomyit + δi + ηit(A.2)

when we follow a panel fixed-effect approach (Equation (2)). Results are in columns 4 and 5.

∆ln(Lane kmit) = γ1 × ∆ln(Historical (rail)road kmit)

+ γ2 × ∆ln(Populationit) + γ3 × ∆Socioeconomyit + ∆ηit(A.3)

when we follow a first-difference approach (Equation (3)). Results are in columns 6 to 11.

Table A.1: Historical (rail)roads and modern highway lane kilometers, OLS results: First-stage

Dependent variable: ln(Lane km) ∆ln(Lane km)

Pool Panel First-Difference

[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11]

ln(Roman roads) 0.026a 0.025a 0.025a 0.037a 0.038a ∆ln(Roman roads) 0.034a 0.034a 0.028a 0.028a 0.034a 0.034a

(0.007) (0.007) (0.007) (0.009) (0.009) (0.007) (0.007) (0.007) (0.007) (0.008) (0.008)

ln(15th c. trade routes) 0.002(0.007)

ln(1810 post routes) 0.011(0.008)

ln(1870 railroads) 0.029a 0.027a 0.008 ∆ln(1870 railroads) 0.006 0.002 0.011(0.008) (0.008) (0.008) (0.008) (0.008) (0.011)

ln(Population) X X X X X ∆ln(Population) X X X X X XSocioeconomy X X X X X ∆Socioeconomy X X X X X XGeography X X X Geography X XHistory X X X History X XCountry fixed-effects X X X Country fixed-effects X XFUA fixed-effects X X FUA fixed-effects X X

Lagged ln(VKT) X X X XYear fixed-effects X X X X X Year fixed-effects X X X X X X

First-Stage F-Statistic 5.22 10.36 12.01 10.55 19.79 First-Stage F-Statistic 11.13 21.26 7.92 14.12 8.44 16.22

Notes: 1,635 observations (545 cities × 3 decades (1985-2005)) in pooled and panel regressions (columns 1 to 5) and 1,090observations (545 cities × 2 periods) in first-difference regressions (Columns 6 to 11). Geography controls include by the logarithmof the FUA area, a suburbanization index, which is the share of the central city area, the mean and range of FUA elevation, themean surface ruggedness for each FUA and the logarithm of the distance to the closest coast from the central city centroid. Historyincludes the logarithms of FUA population in 1960, 1970 and 1980, and dummy variables for historical major cities in 814, 1000,1200, 1450 and 1850. Socioeconomic characteristics are the log of the GDP, the share of employment in manufacturing, the shareof employment in finance and business services, the share of employment in non-market services, and the unemployment rate.Robust standard errors are clustered by FUA and are in parenthesis. a, b and c indicates significant at 1, 5, and 10 percent level,respectively.

22

A.3 First-stage results for highway length and number of lanes

Table A.2: Historical (rail)roads and modern highway length and lanes, OLS results: First-stage

Dependent variable: ln(Km of highways) Number of lanes

[1] [2] [3] [4]

ln(Km of Roman roads) 0.043a 0.050a Number of all historical (rail)roads 0.119a

(0.015) (0.015) (0.023)

ln(Km of 15th c. trade routes) -0.017b Number of Roman roads 0.183a

(0.007) (0.040)

ln(Km of 1810 post routes) 0.011 Number of 15th c. trade routes 0.174a

(0.010) (0.046)

ln(Km of 1870 railroads) 0.010 Number of 1810 post routes 0.097b

(0.010) (0.049)

Number of 1870 railroads 0.157a

(0.046)ln(Population) X X ln(Population) X XSocioeconomy X X Socioeconomy X XFUA fixed-effects X X FUA fixed-effects X XYear fixed-effects X X Year fixed-effects X X

First-Stage F-Statistic 5.15 11.52 First-Stage F-Statistic 26.29 11.78


A.4 First-stage results for modern railroads

Table A.3: Historical (rail)roads and modern railroads, OLS results: First-stage

Dependent variable: ln(Kilometers of railroads)

[1] [2]

ln(Kilometers of Roman roads) -0.019(0.012)

ln(Kilometers of 15th c. trade routes) 0.006(0.013)

ln(Kilometers of 1810 post routes) 0.011(0.020)

ln(Kilometers of 1870 railroads) 0.308a 0.311a

(0.083) (0.075)ln(Population) X XSocioeconomy X XFUA fixed-effects X XYear fixed-effects X X

First-Stage F-statistic 5.75 17.41


23

Appendix B. Additional results

Table B.1: The effect of highways on traffic congestion, OLS and IV results: Years


OLS results IV Results

1985 1995 2005 1985 1995 2005

[1] [2] [3] [4] [5] [6]

ln(Lane km) 0.904a 0.813a 0.768a 1.832a 1.310a 1.844a

(0.043) (0.039) (0.040) (0.342) (0.225) (0.264)Population X X X X X XGeography X X X X X XHistory X X X X X XSocioeconomy X X X X X XCountry fixed-effects X X X X X X

Adjusted R2 0.88 0.90 0.88First-Stage F-statistic 11.92 11.89 14.11Instrument ln(Km of Roman roads)

Notes: 545 observations in each regression. Geography controls include the logarithm of the FUA area, a suburbanization index,which is the share of the central city area, the mean and range of FUA elevation, the mean surface ruggedness for each FUA and thelogarithm of the distance to the closest coast from the central city centroid. History includes the logarithms of FUA populationin 1960, 1970 and 1980, and dummy variables for historical major cities in 814, 1000, 1200, 1450 and 1850. Socioeconomiccharacteristics are the log of the GDP, the share of employment in manufacturing, the share of employment in finance andbusiness services, the share of employment in non-market services, and the unemployment rate. Robust standard errors areclustered by FUA and are in parenthesis. a, b and c indicates significant at 1, 5, and 10 percent level, respectively.

Table B.2: The effect of highways on traffic congestion, OLS and IV results: Gradual


OLS results IV Results

[1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

ln(Lane km) 0.969a 0.752a 0.896a 0.882a 0.831a 1.480a 1.617a 1.519a 1.529a 1.558a

(0.026) (0.031) (0.033) (0.033) (0.034) (0.150) (0.313) (0.233) (0.256) (0.272)Population X X X X X X X XGeography X X X X X XHistory X X X XSocioeconomy X XCountry fixed-effects X X

Adjusted R2 0.797 0.830 0.852 0.855 0.887First-Stage F-statistic 29.51 13.04 15.89 12.82 12.01Instrument ln(Km of Roman roads)

Notes: 1,635 observations (545 cities × 3 decades (1985-2005)) in each regression. Geography controls include the logarithm of theFUA area, a suburbanization index, which is the share of the central city area, the mean and range of FUA elevation, the meansurface ruggedness for each FUA and the logarithm of the distance to the closest coast from the central city centroid. Historyincludes the logarithms of FUA population in 1960, 1970 and 1980, and dummy variables for historical major cities in 814, 1000,1200, 1450 and 1850. Socioeconomic characteristics are the log of the GDP, the share of employment in manufacturing, the shareof employment in finance and business services, the share of employment in non-market services, and the unemployment rate.Robust standard errors are clustered by FUA and are in parenthesis. a, b and c indicates significant at 1, 5, and 10 percent level,respectively.

24

Appendix C. Robustness checks

Tabl

eC

.1:T

heef

fect

ofhi

ghw

ays

ontr

affic

cong

esti

on:R

obus

tnes

sch

ecks

Dep

ende

ntva

riab

le:

ln(V

KT)

Rob

ustn

ess:

Pane

lA:

Pane

lB:

Pane

lC:

Pane

lD:

Inst

rum

enti

ngpo

pLi

near

tren

dsO

ther

road

sN

on-l

inea

rity

IVIV

OLS

IVO

LSIV

OLS

IV[1

][2

][3

][4

][5

][6

][7

][8

]

ln(L

ane

km)

0.72

3a2.

312a

ln(L

ane

km)

0.69

5a1.

206a

ln(L

ane

km)

0.68

6a1.

444a

ln(L

ane

km)

0.71

3a1.

230a

(0.0

62)

(0.5

34)

(0.0

97)

(0.3

14)

(0.0

98)

(0.3

94)

(0.0

33)

(0.1

65)

ln(K

mse

cond

.roa

ds)

-0.0

21b

-0.0

29b

(ln(

Lane

km))

2-0

.029

-0.2

62(0

.010

)(0

.011

)(0

.020

)(0

.21)

ln(K

mlo

calr

oads

)0.

013

-0.0

12(0

.011

)(0

.020

)ln

(Pop

ulat

ion)

XX

ln(P

opul

atio

n)X

Xln

(Pop

ulat

ion)

XX

ln(P

opul

atio

n)X

X

Soci

oeco

nom

yX

XSo

cioe

cono

my

XX

Soci

oeco

nom

yX

XSo

cioe

cono

my

XX

Geo

grap

hyX

XFU

Afix

ed-e

ffec

tsX

XFU

Afix

ed-e

ffec

tsX

XFU

Afix

ed-e

ffec

tsX

X

His

tory

XX

Year

fixed

-eff

ects

XX

Year

fixed

-eff

ects

XX

Year

fixed

-eff

ects

XX

Cou

ntry

fixed

-eff

.X

XC

ount

rytr

end

XX

Year

fixed

-eff

ects

XX

Adj

uste

dR

2A

djus

ted

R2

0.71

Adj

uste

dR

20.

71A

djus

ted

R2

0.83

Firs

t-St

age

F-st

at12

.06

4.66

Firs

t-St

age

F-st

at19

.79

Firs

t-St

age

F-st

at5.

30Fi

rst-

Stag

eF-

stat

3.71

Inst

rum

ents

ln(R

oman

)In

stru

men

tsln

(Rom

an)

Inst

rum

ents

ln(R

oman

)In

stru

men

tsln

(Rom

an)

Tem

pera

ture

,Pre

cipi

tati

onln

(181

0Po

st),

ln(1

5th

c.Tr

ade)

ln(R

om)×

ln(R

om)

Not

es:

1,63

5ob

serv

atio

ns(5

45ci

ties×

3de

cade

s(1

985-

2005

))in

each

regr

essi

on.

Geo

grap

hyco

ntro

lsin

clud

eth

elo

gari

thm

ofth

eFU

Aar

ea,a

subu

rban

izat

ion

inde

x,w

hich

isth

esh

are

ofth

ece

ntra

lcit

yar

ea,t

hem

ean

and

rang

eof

FUA

elev

atio

n,th

em

ean

surf

ace

rugg

edne

ssfo

rea

chFU

Aan

dth

elo

gari

thm

ofth

edi

stan

ceto

the

clos

est

coas

tfr

omth

ece

ntra

lcit

yce

ntro

id.

His

tory

incl

udes

the

loga

rith

ms

ofFU

Apo

pula

tion

in19

60,1

970

and

1980

,and

dum

my

vari

able

sfo

rhi

stor

ical

maj

orci

ties

in81

4,10

00,1

200,

1450

and

1850

.So

cioe

cono

mic

char

acte

rist

ics

are

the

log

ofth

eG

DP,

the

shar

eof

empl

oym

ent

inm

anuf

actu

ring

,th

esh

are

ofem

ploy

men

tin

finan

cean

dbu

sine

ssse

rvic

es,

the

shar

eof

empl

oym

ent

inno

n-m

arke

tse

rvic

es,

and

the

unem

ploy

men

tra

te.R

obus

tst

anda

rder

rors

are

clus

tere

dby

FUA

and

are

inpa

rent

hesi

s.a ,

ban

dc

indi

cate

ssi

gnifi

cant

at1,

5,an

d10

perc

ent

leve

l,re

spec

tive

ly.

25

Appendix D. An alternative approach to analyze the role of tolls

As an alternative, but also complementary approach to the one developed in Section 5.1 to analyzethe role of tolls, we consider an empirical strategy based on subsamples. Specifically, in Table D.1we split our sample of cities and run separate regressions. First, in columns 1 and 2 we comparecities without tolls (285) with cities with tolls (260). Second, in columns 3 and 4 we compare citieswith less than 25% of tolled highways (343) with cities with 25% or more of tolled highways (202).All LIML results point out in the same direction: The effect of lane kilometers on VKT is higherin cities without tolls or with less than 25% of tolled highways than in cities with tolls or withmore than 25% of tolled highways. Furthermore, LIML results show that the fundamental law isonly related to cities without tolls or with less than 25% of tolled highways, with elasticities of1.13 and 1.34, respectively. As a whole, these results confirm the ones obtained using an empiricalstrategy based on interactions (Section 5.1).

Table D.1: The effect of highways on traffic congestion, IV results: Tolls and subsamples


Without vs. With tolls Below vs. Above 25% tolled highways

without with below above[1] [2] [3] [4]

ln(Lane km) 1.132c 0.856a 1.335b 0.589c

(0.680) (0.328) (0.529) (0.328)ln(Population) X X X XSocioeconomy X X X XFUA fixed-effects X X X XYear fixed-effects X X X X

Observations 855 780 1029 606FUA 285 260 343 202First-Stage F-statistic 4.82 11.16 7.72 9.44Instrument ln(Km of Roman roads) ln(Km of Roman roads) ln(Km of Roman roads) ln(Km of Roman roads)

Notes: Socioeconomic characteristics are the log of the GDP, the share of employment in manufacturing, the share of employmentin finance and business services, the share of employment in non-market services, and the unemployment rate. Robust standarderrors are clustered by FUA and are in parenthesis. a, b and c indicates significant at 1, 5, and 10 percent level, respectively.

26

Núm Títol Autor Data

20.10 Forgetting –by-not-doing: The case of surgeons

and cesarean sections

Gabriel A. Facchini Palma

Octubre

2020

20.09 Low Staffing in the Maternity Word: Keep Calm and

call the Surgeon

Gabriel A. Facchini Palma

Setembre

2020

20.08 Impact of public transport strikes on air pollution

and transport modes substitution in Barcelona

Lyna González / Jordi

Perdiguero / Àlex González

Setembre

2020

20.07 Globalización y responsabilidad en los problemas

ecológicos

Jordi Roca Jusmet /

Emilio Padilla Rosa

Setembre

2020

20.06 Do short-term rental platforms affect housing

markets? Evidence form Airbnb in Barcelona

Miquel-Àgel Garcia / Jordi

Jofre / Rodrigo Martínez /

Mariona Segú

Setembre

2020

20.05 The driving factors of CO2 emissions from

electricity generation in Spain: A decomposition

analysis

Vicent Alcántara /

Emilio Padilla /

Pablo del Río

Juliol 2020

20.04 CO2 What do divided cities have in common?

Am international comparison of income

segregation

Paolo Veneri/

Andre Comandon /

Miquel-Àngel Garcia López /

Michiel N.Daams

Juliol 2020

20.03 CO2 emissions of the construction sector in Spain

during the real estate boom: input-output

subsystem analysis and decomposition

Vicent Alcántara/

Emilio Padilla

Juny 2020

20.02

The Direct Rebound Effect of Electricity Energy

Services in Spanish Households: Evidence from

Error Correction Model and System GMM

estimates

Martín Bordón /

Jaume Freire /

Emilio Padilla

Maig 2020

20.01 Subsidizing Innovation Over the Business Cycle Isabel Busom / Jorge-Andrés

Vélez-Ospina

Març 2020

19.07 África Subsahariana: ¿Del afropesimismo a la

transformación económica?

Artur Colom Jaén Novembre

2019

19.06 Identification of relevant sectors in CO2 emissions

in Ecuador through input-output analysis

Edwin Buenaño / Emilio

Padilla and Vicent Alcántara

Setembre

2019

19.05 Driving forces of CO2 emissions and energy

intensity in Colombia

Lourdes Isabel Patiño / Vicent

Alcàntara and Emilio Padilla

Setembre

2019

1 Introduction2 Congestion and highways in Europe2.1 Congestion in Europe2.2 Highways in Europe

3 Empirical strategy3.1 Pool, panel and first-difference3.2 Instrumental variables

4 Does the fundamental law of highway congestion apply in Europe?4.1 Main results4.2 Coverage and capacity effects

5 Do tolls and public transportation matter?5.1 Tolls5.2 Public transportation

6 ConclusionsReferencesAppendix A Historical roads and railroads in EuropeA.1 MapsA.2 First-stage results for highway lane kilometersA.3 First-stage results for highway length and number of lanesA.4 First-stage results for modern railroads

Appendix B Additional resultsAppendix C Robustness checksAppendix D An alternative approach to analyze the role of tolls

Documents

Departament d´Economia Aplicada · Universitat de Barcelona and Institut d’Economia de Barcelona October 2020 Abstract: Using data from the 545 largest European cities, we study